This doesn’t really work though. If it were simply directing you to actual sources that could help out (articles, and lawyer authored blogs or something), that would be one thing, but LLMs are just probability machines. It’ll write out what looks like good sound reasoning and logic, but in reality it’s simply the most probable word followed by the next most probable word, over and over again. It’s using pattern matching, not knowledge
They output responses that read like fact checks, and link things that are semantically similar to their output.
LLMs don’t understand concepts like “fact” and “source” though, so it’s a stretch to say they “do fact checks” and “provide online sources”.
Always remember, LLMs don’t actually understand anything other than what order words typically go in. They’re really good at it, sure, but don’t let that trick you into thinking they have any deeper understanding than that.
I’ve tried on several occasions to try and explain this to people, and it’s like they WANT it to be a sentient thing, and refuse all evidence to the otherwise.
They do say that. I’ve seen Claude realise some programming package must be internal / non public and it therefore can’t provide any information on that.
It’s nothing more than surface level though. Their “knowledge” is limited to whatever is present in their training data, and any claims it makes outside of that are manufactured based on statistics and it’s knowledge of word order, not actual understanding of actual facts.
Just because it gets it right sometimes doesn’t mean it knows it’s right. Because, again, the only thing LLMs actually understand is word order.
No it’s not limited to their current training data. Like I said, they can search the internet, get information from websites or available studies, summarize and use that in their answers.
Just because it gets it right sometimes doesn’t mean it knows it’s right.
This doesn’t really work though. If it were simply directing you to actual sources that could help out (articles, and lawyer authored blogs or something), that would be one thing, but LLMs are just probability machines. It’ll write out what looks like good sound reasoning and logic, but in reality it’s simply the most probable word followed by the next most probable word, over and over again. It’s using pattern matching, not knowledge
Have you used the newest LLM models? They certainly do fact checks, research info and provide online sources - if they are instructed to do that.
They output responses that read like fact checks, and link things that are semantically similar to their output.
LLMs don’t understand concepts like “fact” and “source” though, so it’s a stretch to say they “do fact checks” and “provide online sources”.
Always remember, LLMs don’t actually understand anything other than what order words typically go in. They’re really good at it, sure, but don’t let that trick you into thinking they have any deeper understanding than that.
I’ve tried on several occasions to try and explain this to people, and it’s like they WANT it to be a sentient thing, and refuse all evidence to the otherwise.
So basically they do what humans do. What exactly qualifies as deeper understanding to you?
Being able to say “I don’t know” rather than making shit up to fill in what it doesn’t know. That would be a good start.
They do say that. I’ve seen Claude realise some programming package must be internal / non public and it therefore can’t provide any information on that.
It’s nothing more than surface level though. Their “knowledge” is limited to whatever is present in their training data, and any claims it makes outside of that are manufactured based on statistics and it’s knowledge of word order, not actual understanding of actual facts.
Just because it gets it right sometimes doesn’t mean it knows it’s right. Because, again, the only thing LLMs actually understand is word order.
No it’s not limited to their current training data. Like I said, they can search the internet, get information from websites or available studies, summarize and use that in their answers.
Again, how do you know?
Because I’ve studied LLMs and the underlying concepts they’re based on.
deleted by creator
Do you?
Yes
How do you know that you don’t just guess words, too?